Skip to main content
European Commission logo print header

Variational Basis Learning for Statistical Motion Atlases: Application to Quantitative Dynamic Cardiac Imaging

Final Report Summary - BALMORAL (Variational Basis Learning for Statistical Motion Atlases: Application to Quantitative Dynamic Cardiac Imaging)

Pulmonary Arterial Hypertension (PAH), is a severe progressive disorder characterised by a vasculopathy of the small pulmonary arteries (<500 μm diameter) to the lung. Failure of the right ventricle (RV) to adapt to elevated resistance to blood flow (afterload) results in death, usually within 3 years. Right heart catheterisation (RHC) is used to diagnose and, in selected patients, to assess response to treatment. RHC is an invasive test, which provides indirect information relating to right ventricular function and does not directly probe the right ventricle. Current therapies for PAH are very expensive (up to €28,000 per patient per annum) and there is an urgent need for specific markers to assess treatment response and for patient risk stratification. Image-based global measures such as RV volume can only reflect the overall performance of the RV. However, there is good evidence that PAH can be identified by morphological and motion abnormalities in the RV and the interventricular septum (IVS) without the need for invasive and expensive RHC.

One specific clinical target of the proposed research plan was to investigate whether we can quantify PAH in severity using imaging features that are extracted from the myocardium in IVS or RV. From a clinical standpoint, to the best of our knowledge, this proposal was the first attempt to develop computational diagnostic biomarkers of the right heart. Nevertheless, from a technical point of view, we extended beyond the state of the art and propose a highly novel and advanced probabilistic method for the discovery of abnormalities adaptable to other diseases, such as Hypertrophic Cardiomyopathy (HCM), and organs.

The initial imaging biomarker or the pattern of interest considered in this project was the motion of the heart, however, we also realized that the morphology (shape) of the ventricles have a significant diagnostic value. Given the fact the motion can be described as a variable sequence of shape patterns over time, our framework was designed to be flexible to handle both traits. Inspired by this observation, we assumed that in our patient population the pattern of interest in each patient is described as a point set containing high-dimensional points. For instance, for shape analysis points are three dimensional, representing the boundary the object. As such, for motion analysis points can be made by concatenating the subsequent 3D positions of the landmarks on the boundary of the heart, resulting in a high-dimensional description. Although, this assumption generalizes well, there were certain complexities associated with our modelling; for instance, the number of points in each set can be different, thus point-to-point correspondence from one patient to another, and hence consistent vector presentation of the data across the population was inaccessible.

Another challenge in the project was due to uncertainty in the diagnosis of the patients, who do not fully develop disease related abnormalities. As a result, in a fully supervised scenario, where all the patients are labelled as either healthy or otherwise by clinical experts, these subtle differences indicating the severity or subtypes of the disease are often ignored. Thus, the discovery of new disease subtypes is often hampered and limited to the existing clinical knowledge.

To address these particular challenges arisen from the data, we designed an unsupervised machine learning framework that significantly advances the available tools and addresses the aforementioned challenges through a fundamentally new approach. In this project, we presented a generative model to infer the pdf of unstructured, rigidly aligned point sets having no point-to-point correspondences. The framework is a piece-wise linear model for join clustering of point sets, and estimating the local modes of variations in each cluster. Points at each set are regarded as samples from a low dimensional Gaussian Mixture Model (GMM), whose means are concatenated to form higher dimensional vectors. These vectors are considered as samples from a Mixture of Probabilistic Principal Component Analyzers (PPCA). The latter is a high dimensional GMM, whereas the covariance matrices of its clusters are explicitly decomposed to subspaces of local principal as well as random (isotropic) variations. An inference algorithm based on variational Bayes (VB) is proposed for unsupervised learning of class labels and variations.

In summary, in line with the outlined objectives in the project, the following contributions and achievements were made:
• Using mixture of PPCA, a larger class of pdfs was modelled, leading to more realistic group means and local variation modes. Unlike the available methods, variation modes were explicitly modelled, eliminating the post PCA step. Moreover, we handled point sets having no point-to-point correspondences and derived a lemma for shape prediction and projection to the space spanned by the local PPCA's.
• We proposed a full Bayesian model and provided an explicit tight lower bound on the model evidence given data. By maximizing the later, discrete parameters such as numbers of clusters and variation modes (basis) were determined, enabling automatic model selection.
• The model is highly flexible; different and successful applications of the framework were demonstrated to joint clustering and component analysis of population data from healthy, PHA and HCM cardiac morphologies, spatio-temporal patterns from patients with cardiac strokes versus normal hearts, lumbar vertebrae. The results are being disseminated through publications in high-rank international journals and conferences.
• Thanks to the versatility of the framework, other novel applications and thus a larger impact on the field is acknowledged. For instance, we extended the framework to predict the wall-shear stress on intracranial aneurisms, given their morphologies. We are planning to utilize this model to construct atlases of coronary arteries. Our tool is available for download through the home page of the host center (www.cistib.org).

Apart from research and technological advancements, this Marie-Curie fellowship allowed the researcher to develop a significant set of leadership skills and establish his own career independence. By the end of this project, the researcher was promoted as a Lecturer in Medical Image Computing in University of Sheffield.